A preliminary study for EEG-based BCI development for FES control

The use of motor imagery to activate motor-related brain areas represents an effective tool for promoting motor rehabilitation in individuals with severe muscle deficits. The oscillations in neuronal activity occuring during motor imagery can be acquired and processed into external outputs by using Brain Computer Interface (BCI) systems. To improve the effectiveness of rehabilitation procedures, it is useful to combine BCI and a Functional Electrical Stimulation (FES) device, which takes as input the signal processed by the BCI and translates it into a stimulation that generates movement in the compromised limb. Indeed, an approach based on simultaneous activation of cortical regions (through motor imagery) and motor nerves of the muscle of interest (through FES) has the potential to promote functional reorganization of cortical structures, thereby improving the results of the rehabilitation procedure. Therefore, the development of new techniques for acquiring and processing neural activity to provide accurate inputs for FES activation plays a central role in the field of rehabilitation. In this thesis work, two algorithms suitable for non-invasive BCIs applications were designed. The simplest signal recording method for BCI systems is ElectroEncephaloGraphy (EEG). EEG is a non-invasive technique that provides useful information to recognize the subject¿s motion intention, which is the event of interest for the activation of a FES device. The proposed algorithms process EEG signals recorded during motor imagery tasks, which are characterized by changes in cortical fluctuations in a specific frequency band, typical of the mu/alpha brain rhythm. These changes were detected by a threshold, iteratively determined and subject-specific, that defines the trigger event of motor intention. The first algorithm was developed with a standard approach and performs a signal analysis in the frequency domain. The parameter used in this procedure is the Power Spectral Density (PSD) of the signal, which shows amplitude suppression during motor imagination. The second algorithm was developed using a non-linear operator defined in the continuous domain, called Teager¿s Energy Operator (TEO), which tracks energy changes in the signal. Besides the threshold, in this approach an amplitude constraint was introduced to minimize the influence of artifacts and increase the accuracy of event recognition. The goal of both implemented techniques was a fast, real-time detection of the motor intention events, and two different approaches were tested and compared, to assess which one provided the best performance in terms of sensitivity, specificity, accuracy, precision and activation delay. The tests were performed on two datasets, one provided by BCI competiton IV and the other consisting of signals acquired on-site with g.tec recording system. To minimize processing complexity and enable the development of a low-power and wearable system, the use of a single EEG channel was investigated. Application of the proposed methods to the on-site recorded dataset led to online accuracy rates of 83.7% for the PSD-based algorithm and 84.5% for the TEO-based one, while the movement onset detection latency was 240 ms and 771 ms for the two approaches, respectively.